What RFID badge tracking at exhibitions actually measures and what it misses

An exhibitor in hall two asked a fair question after the show. Of the people who came to my stand, where did they go next? The organiser had spent a considerable sum on RFID badge tracking at exhibitions that year, so the answer felt like it should be sitting in a table somewhere.

It was not, and the reason has nothing to do with the vendor. Passive badge tracking observes people at the places you put readers. Between those places it observes nothing at all, and the gaps are where most of the hall is.

What does an RFID portal actually record?

A passive UHF badge has no battery. The reader illuminates it, the tag backscatters its identifier, and the reader logs a read. ISO/IEC 18000-63 specifies that air interface at 860 to 960 MHz, defining passive backscatter, interrogator-talks-first behaviour for readers and tags, with published editions in 2013, 2015 and 2021.

What lands in your database is one row per read: tag identifier, reader identifier, timestamp, sometimes a signal strength value. A portal is a pair of antennas either side of an aisle or doorway, and a crossing is a burst of reads collapsed into one event.

So the record of an attendee's day is a sequence of crossings. Portal 4 at 10:12, portal 9 at 10:31, portal 9 again at 11:05, portal 2 at 11:40. Everything else is reconstruction. If your hall has twelve portals across 9,000 square metres, you have twelve observation points and no information about the space between them.

Chongwatpol used exactly this design in the European Journal of Operational Research in 2015, distributing RFID badges to build a traceability framework for a trade show and then asking which stands attracted the most interest, how long visitors spent at each, and how many times they returned before buying. That paper is the clearest published statement of what the technology supports at an exhibition, and its questions are all about counts and durations at known points.

Portal placement is a budget decision dressed up as a design decision. Each portal has a cost in hardware, cabling, power and a floor plan argument with the operations team about where a gantry can stand, so nobody installs one per aisle. They go at hall entrances, at the ends of main aisles, and at feature areas somebody wanted measured. That placement then decides which questions are answerable for the next five years, because the analysis you can run is bounded by where the antennas went, and a question that occurs to the commercial team in March cannot be answered from data collected in September.

Read rate, and what it does to a journey

Every portal has a read rate below one. The number that matters is not the rate itself but what happens when you chain several crossings together, because a journey survives only if every crossing in it was read.

Take a portal at 96 per cent. A two-crossing journey survives at 0.96 squared, which is 0.9216, so 7.8 per cent of them are broken. A four-crossing journey survives at 0.96 to the fourth, which is 0.84935, so 15.1 per cent are broken. A six-crossing journey survives at 0.78276, so 21.7 per cent are broken.

Now drop the portal to 90 per cent, which is entirely realistic on a wide entrance with crowds arriving in clumps. The four-crossing journey survives at 0.9 to the fourth, which is 0.6561. A third of your journeys are now incomplete.

Put that against a real population. 8,400 badges in the hall, each making four crossings, at 96 per cent per portal: 8,400 times 0.84935 gives 7,134 complete sequences and 1,266 broken ones. At 90 per cent it gives 5,511 complete and 2,889 broken.

The counts hold up far better than the sequences. Total crossings at portal 4 are wrong by 4 per cent at a 96 per cent read rate, which is a bias you can state and live with. The set of people whose full route you can reconstruct is wrong by 15 per cent, and it is wrong non-randomly, since the people who move most have the most chances to be missed. Any analysis of the busiest attendees is running on the population least likely to be complete.

Three things that break a read

Badge orientation comes first. A UHF tag is a dipole, and its read range collapses when the tag sits perpendicular to the antenna polarisation. A badge on a lanyard rotates freely, so the same person on the same walk through the same portal presents a different orientation every time.

Bodies come second. Water absorbs at UHF frequencies and a human torso is mostly water. A badge worn against the chest of somebody walking with their back to the antenna is being read through the worst possible medium. This is why read rates fall exactly when the hall is busiest, which is exactly when the data matters most.

Tag collision comes third. When twenty badges enter the field together, the reader has to run an anticollision procedure to separate them, and the time available is bounded by how fast people are walking. Dense arrivals lose more tags than sparse ones, again in the peak.

All three failures correlate with crowding. That is the property to hold on to, because it means your missing data is concentrated in the busy periods, and a naive comparison of a quiet Wednesday against a busy Tuesday will understate Tuesday.

There is a cheap way to watch read rate continuously once the portals are up. Fix a small number of reference tags at known positions inside each portal field and record whether the reader sees each of them in every polling cycle. A reference tag that goes quiet tells you the portal has degraded, in minutes rather than in the debrief, and the pattern across portals separates a hardware fault from a hall that has simply filled up. The reference tags cost almost nothing next to the portals themselves and most installers will fit them if you ask during commissioning.

Derived dwell and the exit read you never got

Portal systems produce dwell by subtraction. A badge crosses into a zone at 13:02 and out at 13:49, so the dwell is 47 minutes. Clean when both reads exist.

Miss the exit read and the calculation has nowhere to stop. The zone stays open until the next crossing anywhere, which might be at 15:20 in a different hall, and the 47 minute dwell becomes 138 minutes. One missing read has produced a figure that is wrong by a factor of three, and it will sit in a distribution alongside genuine values with nothing marking it out.

Two defences work. Cap the open interval at a plausible maximum and flag anything that hits the cap as censored instead of dropping it, so the count of censored intervals becomes a data quality metric you can watch. Second, compute dwell only for badges whose entry and exit are both present, and publish the share of intervals that qualified. If 78 per cent of intervals had both reads, say so next to the median dwell figure.

Comparing what portal dwell measures against the proximity and trajectory approaches is a separate argument, worked through in dwell time measurement. The reading of timestamps inside an exhibitor's own scan file, which has different failure modes again, sits with badge scan timestamps.

What can you honestly tell an exhibitor?

Quite a lot, provided you stay inside what the portals saw.

You can report crossings at each portal by 15 minute bin, with a stated read rate. You can report how many distinct badges crossed a zone boundary, and how that compares with the same zone last edition on the same portal hardware. You can report the distribution of complete-interval dwell times, with the completeness share attached. You can report return crossings, meaning badges that crossed the same portal more than once with a gap long enough to be a genuine second visit.

You cannot report a path, a heat map of the whole hall, or the phrase footfall past your stand unless there is a portal at the stand. You cannot report attention, because a crossing says a badge moved through a doorway.

The temptation to fill the gaps with interpolation is strong and I would resist it. An inferred route between portal 4 and portal 9 assumes people walk shortest paths, and the entire reason exhibitors buy stands near the coffee is that people do not. Where the portals cannot see, the honest output is a stated absence, and the counting coverage argument in gate coverage applies to interior portals exactly as it applies to doors.

Where this stops

Read rate is measurable and almost nobody measures it. Every number above depends on knowing your own, and a vendor's laboratory figure will not tell you what a wet Tuesday in hall two produced.

The measurement itself takes an hour. Stand at a portal with a manual counter, count every person who crosses, and compare against the portal's own count for the same window. Do it twice, once in the first hour and once at the busiest point, because the difference between those two numbers is the thing that distorts your comparisons across the day.

There is also a limit that no measurement fixes. Passive tracking tells you where badges went, and a badge in a bag, clipped to a coat over an arm, or left on a table in the catering area, is still generating a record of nothing. Every portal system carries a tail of badges with implausible movement patterns, and cleaning that tail is judgement rather than arithmetic. Reporting it as part of the attendee analytics method, alongside the read rate, is what keeps the rest of the numbers credible.

Start this week by asking whichever supplier ran your portals for the raw read log rather than the summary. Count reads per crossing at one portal. If the ratio is below about three, your antennas are underpowered for the aisle width and everything downstream is inheriting it.